Robust Fake News Detection under Adversarial Social Context
Social responses are an important input for comment-aware fake news detection. They also create an attack surface: malicious comments can manipulate the context used by a detector. Research on this problem connects adversarial text generation, mechanism-specific evaluation, and adaptive robust training.
Measuring vulnerability across mechanisms
Aggregate detection performance summarizes the overall result, while group-wise measurements reveal how the detector responds to different attacks. Evaluating several comment settings makes this view useful for studying the effect of both attack mechanism and context composition.
The CogniDir experiments use Weibo16, Weibo20, and RumourEval-19. The project page presents detection results, attack success rates, and ablations of malicious comment generation, vulnerability scoring, and allocation.
Adapting the training mixture
Adaptive training allocates exposure according to the detector's current vulnerabilities. In CogniDir, group-wise accuracy and probabilistic confidence inform vulnerability scores. InfoDirichlet Resampling maps these scores to Dirichlet parameters and uses the mean to set the next training mixture.
This feedback connects evaluation with training: the detector is assessed across attack mechanisms, the allocation is updated, and subsequent training responds to the measured weaknesses.
Connecting generation, sampling, and defense
Generate First, Then Sample studies LLM-generated fake news and reinforcement-learning-based sampling of real and fake training examples. SINCon studies malicious message injection in rumor detection and uses contrastive learning to make node influence more uniform in message propagation trees.
CogniDir contributes mechanism-specific malicious comments and feedback-driven allocation across attack groups. These methods provide useful reference points for researchers studying generated training data, adversarial social context, and adaptive robustness. The linked project page and repository collect the experiments, training entry points, and official EMNLP 2026 citation.
Studying adversarial social context
A social-context evaluation records the news item, the comments supplied to the detector, and the resulting prediction. LLMs make it possible to generate comments with different rhetorical and cognitive mechanisms, creating structured tests of detector robustness.
CogniDir studies fact distortion, logical confusion, and emotional manipulation as distinct attack groups. This organization connects the design of malicious comments with group-wise vulnerability measurements.